GRASP: Graph-Reasoning Aided Survey Planning

• The paper presents GRASP that combines LLM planning with two-layer graph representations to capture intra- and inter-paper relationships for high-fidelity related work generation.
• It employs topology-aware pruning using a Steiner tree algorithm to retain critical consensus nodes and central thought units, ensuring structural fidelity.
• The evaluation demonstrates that GRASP (pruned) outperforms baselines by effectively balancing discourse roles, citation importance, and ordering in survey generation.

Searching arXiv for the GRASP paper to ground the article in the original preprint. GRASP, short for Graph-Reasoning Aided Survey Planning, is a framework for related work generation that combines LLM planning with graph algorithms in order to model how cited papers “build on, challenge, or offer alternative perspectives to one another” (Li et al., 4 Jul 2026). The method is presented in “GRASP: Graph-Reasoning Aided Survey Planning for High-Fidelity Related Work Generation,” published on 2026-07-04, and is designed specifically for writing related work sections that preserve citation discourse structure rather than merely maximizing lexical overlap (Li et al., 4 Jul 2026). Its core design consists of a two-layer graph structure—a Graph of Thoughts and an Argument-Counterargument Planning Network—together with topology-aware pruning via a Steiner tree, followed by a staged drafting pipeline that serializes graph structure into JSON for LLM-guided generation (Li et al., 4 Jul 2026).

1. Definition and Scope

GRASP is defined as a framework “combining LLM planning for related work generation with graph algorithms to extract key relationships among cited papers” (Li et al., 4 Jul 2026). The motivating premise is that literature review writing requires more than identifying individually relevant papers: it requires explicit representation of inter-paper relations, including support, contrast, and alternative positioning. In the GRASP formulation, these relations are not treated as incidental rhetorical artifacts but as objects of planning.

The framework targets related work sections, abbreviated in the paper as RWS. Its stated objective is high-fidelity related work generation, with “fidelity” operationalized through citation analysis-based evaluation rather than through traditional overlap metrics alone (Li et al., 4 Jul 2026). Concretely, the paper reports that GRASP generates related work sections that “closely match human-written targets in terms of the discourse roles, intents, and grouping of citations” (Li et al., 4 Jul 2026).

A common misunderstanding would be to regard GRASP as an LLM-only prompting strategy. The paper’s formalism does not support that interpretation. Instead, it specifies a hybrid architecture in which LLM outputs are embedded into graph construction, graph pruning, and graph-conditioned drafting. This suggests that the contribution lies as much in structured planning and topology control as in prompting.

2. Two-Layer Graph Representation

The first layer is the Graph of Thoughts, denoted $G_{\text{GoT}} = (V, E, w)$, defined over a topic cluster of cited papers (Li et al., 4 Jul 2026). The node set $V$ contains “thought” nodes, which may be either original sequential Chain-of-Thought steps or merged consensus nodes. The edge set $E \subseteq V \times V$ contains two edge types: sequential edges connecting successive steps in one paper’s chain of thought, and consensus edges linking an original node to a consensus node or consensus-to-consensus. The weight function $w : E \to \mathbb{R}^+$ is nonnegative; for sequential edges, $w(e)$ may be uniform or based on semantic continuity, while for consensus edges, $w(e)$ can reflect semantic distance or similarity of merged nodes (Li et al., 4 Jul 2026).

Its construction pipeline is defined in four stages. First, the cited set $R$ is partitioned into topics $T_1, \dots, T_k$ by LLM clustering. Second, for each cited paper $r_j \in T_i$, an LLM extracts a Chain-of-Thought $\bigl[t^j_1 \to t^j_2 \to \cdots\bigr]$. Third, the framework creates original nodes $V$0 and sequential edges $V$1. Fourth, node pairs $V$2 are compared with an LLM-based similarity prompt; if they are semantically close, they are merged into a consensus node $V$3, whose text is a summary of all merged originals, and incoming and outgoing edges are rewired to $V$4 (Li et al., 4 Jul 2026).

The second layer is the Argument–Counterargument Planning Network, abbreviated ACPN, represented as a directed, labeled graph

$V$5

where the nodes are the cited papers $V$6, the edge set $V$7 contains paper pairs with a non-neutral relation predicted by an LLM classifier, and the label function satisfies

$V$8

Its construction proceeds by extracting a core claim set $V$9 for each paper $E \subseteq V \times V$0, then feeding $E \subseteq V \times V$1, optionally together with GoT snippets, to an LLM relation-prediction prompt to obtain $E \subseteq V \times V$2, and finally assembling all directed edges into $E \subseteq V \times V$3 (Li et al., 4 Jul 2026).

These two layers operate at different granularities. The paper states that the Graph of Thoughts and the Argument-Counterargument Planning Network represent the cited papers “at different levels of granularity” (Li et al., 4 Jul 2026). The GoT captures intra-topic reasoning structure and cross-paper consensus at the level of extracted thought units, whereas ACPN captures paper-level support and contrast structure. A plausible implication is that the system separates semantic compression from rhetorical positioning: the GoT compresses and organizes local content, while ACPN constrains comparative discourse among papers.

3. Steiner Tree–Based Topology-Aware Pruning

GRASP applies topology-aware pruning to the topic subgraph $E \subseteq V \times V$4 using a Steiner tree formulation (Li et al., 4 Jul 2026). Given a terminal set $E \subseteq V \times V$5, the objective is to find

$E \subseteq V \times V$6

subject to $E \subseteq V \times V$7 being a connected subgraph spanning all $E \subseteq V \times V$8 (Li et al., 4 Jul 2026). The terminal set is defined as all consensus nodes together with the top 20% sequential nodes by betweenness-centrality.

The pruning algorithm is specified as the “2-Approximation Algorithm (Wu & Chao, 2004).” It first computes the metric closure $E \subseteq V \times V$9 on the terminal set $w : E \to \mathbb{R}^+$0 by calculating shortest-path distances in $w : E \to \mathbb{R}^+$1. It then computes a minimum-spanning tree $w : E \to \mathbb{R}^+$2 on $w : E \to \mathbb{R}^+$3, unpacks each MST edge back to the corresponding shortest path in $w : E \to \mathbb{R}^+$4, and outputs the resulting connected subgraph $w : E \to \mathbb{R}^+$5 (Li et al., 4 Jul 2026). The paper states that this yields a connected pruned subgraph $w : E \to \mathbb{R}^+$6 with $w : E \to \mathbb{R}^+$7.

The significance of this step is explicit in the paper’s terminology: the pruning is “topology-aware” because it is not a flat token- or sentence-level compression heuristic, but a graph-theoretic selection process that preserves terminals judged structurally important. Since the terminal set includes all consensus nodes and high-betweenness sequential nodes, the retained structure emphasizes both merged shared ideas and topologically central reasoning steps. This suggests that GRASP attempts to preserve connectivity among salient argumentative units rather than merely selecting individually salient units.

Another possible misconception is that pruning is a post hoc shortening device. In the GRASP design, pruning is part of planning: the pruned topology is the structure that conditions later drafting. The quantitative comparison between “GRASP (unpruned)” and “GRASP (pruned)” further indicates that pruning is not only a compression step but also a fidelity-improving structural intervention (Li et al., 4 Jul 2026).

4. Integration with LLM Planning and Drafting

The LLM integration is described as a pipeline rather than a single prompt. Chain-of-Thought extraction leads to GoT assembly through “prompt #1 in Appendix,” claim extraction prepares ACPN nodes through “prompt #2,” relation classification produces ACPN edges through “prompt #3,” and a writer prompt pipeline, designated “prompt #4,” uses serialized GoT plus ACPN in JSON (Li et al., 4 Jul 2026). The paper therefore treats prompting as stage-specific and graph-conditioned.

Drafting itself is organized into three stages: “Comprehensive Draft (max coverage),” “Semantic Compression (remove redundancy),” and “Final Merging (balance fidelity & brevity)” (Li et al., 4 Jul 2026). These stages indicate that generation is not a direct one-pass realization from graph to prose. Instead, the output is progressively refined with distinct optimization priorities.

The iterative refinement is explicitly guided by graph insights. The GoT topology suggests paragraph breaks; ACPN labels guide comparative or contrastive sentences; and the LLM is prompted to traverse the JSON node-link structure when generating each sentence (Li et al., 4 Jul 2026). In this formulation, paragraph segmentation, citation grouping, and rhetorical contrast are linked to graph structure rather than left entirely to free-form generative behavior.

This architecture clarifies the meaning of “survey planning” in GRASP. Planning is not limited to topic ordering. It includes cluster formation, thought extraction, consensus formation, relation labeling, topology-aware pruning, and graph-conditioned sentence realization. A plausible implication is that the framework can be understood as imposing explicit intermediate representations between document retrieval and final prose generation.

5. Evaluation Methodology

The evaluation is citation-analysis based and extends beyond standard summarization metrics (Li et al., 4 Jul 2026). Four metric families are specified.

First, sentence discourse roles are measured using the CORWA tagger. For each role $w : E \to \mathbb{R}^+$8, the generated ratio

$w : E \to \mathbb{R}^+$9

is compared to the gold ratio $w(e)$0, and the reported quantity is $w(e)$1 (Li et al., 4 Jul 2026). This metric evaluates whether the rhetorical composition of the generated related work matches the target.

Second, citation importance is measured as Dominant versus Reference. Each cited paper is treated as an instance with gold label $w(e)$2, and precision, recall, and F1 are computed (Li et al., 4 Jul 2026). This captures whether the generator correctly identifies which papers are central versus merely background references.

Third, citation intent is measured with the MultiCite tagger. The paper reports multi-class accuracy and per-class F1 on labels in $w(e)$3 (Li et al., 4 Jul 2026). The detailed highlight table includes Background, Differences, Extends, Motivation, Similarities, and Uses.

Fourth, citation co-occurrence is measured in two ways. The first is edge-connected Jaccard: for each paper $w(e)$4, $w(e)$5 is the set of papers co-cited in the same paragraph in the generated RWS, and $w(e)$6 is the corresponding set in the gold RWS; the average Jaccard coefficient is then reported. The second is citation ordering, measured with Kendall’s $w(e)$7 between the sequence of citation marks in generated versus gold text (Li et al., 4 Jul 2026).

This methodology is notable because it evaluates discourse roles, intents, and grouping of citations—the same dimensions identified in the abstract as central to fidelity (Li et al., 4 Jul 2026). A plausible implication is that GRASP is positioned against evaluation regimes in which high lexical overlap could coexist with poor citation structure.

6. Quantitative Results and Illustrative Case

The experiments are reported on the “OARelatedWork test set, 1,350 papers” (Li et al., 4 Jul 2026). On traditional metrics, GRASP (pruned) obtains the best scores among the listed systems: ROUGE-1 $w(e)$8, ROUGE-2 $w(e)$9, ROUGE-L $w(e)$0, and BERTScore $w(e)$1 (Li et al., 4 Jul 2026). The comparison includes SRW, $w(e)$2, Direct (LLM-only), No-graph, GRASP (unpruned), and GRASP (pruned).

On Citation Importance, GRASP (pruned) reports Dominant precision $w(e)$3, recall $w(e)$4, and F1 $w(e)$5, alongside Reference precision $w(e)$6, recall $w(e)$7, and F1 $w(e)$8 (Li et al., 4 Jul 2026). For Citation Intent Macro F1 highlights, GRASP (pruned) reports Background $w(e)$9, Differences $R$0, Extends $R$1, Motivation $R$2, Similarities $R$3, and Uses $R$4 (Li et al., 4 Jul 2026). On Citation Co-Occurrence and Ordering, GRASP (pruned) records Edge Jaccard $R$5 and Kendall’s $R$6 (Li et al., 4 Jul 2026).

These results support a distinction between graph use and graph pruning. “No-graph” already exceeds several weaker baselines on some traditional metrics, but the reported best-performing configuration is “GRASP (pruned)” rather than “GRASP (unpruned)” (Li et al., 4 Jul 2026). This suggests that explicit graph construction and topology-aware pruning are both consequential components of the framework.

The paper also provides a condensed example from the test case “Parameter Sharing Methods for Multilingual Self-Attentional Translation Models” by Sachan & Neubig (2018) (Li et al., 4 Jul 2026). In the gold related work, the first segment is labeled “[Background],” followed by “[Transition + Grouping],” and then “[Reflection]”; the cited works include Ando and Zhang (2005), Collobert et al. (2011), Dong et al. (2015), Firat et al. (2016), Zoph et al. (2016), and Gu et al. (2018) (Li et al., 4 Jul 2026).

The corresponding GRASP (pruned) output is annotated as Paragraph 1 “[Background + Multi-Paper Summary]” and Paragraph 2 “[Contrast + Method Proposal]” (Li et al., 4 Jul 2026). The paper explicitly observes that GRASP groups “the four key prior systems into one paragraph (matching the human grouping),” uses multi-paper summaries as a discourse role, ends with a contrast or reflection tied to the new contribution, cites each system exactly once, and orders them by chronological and conceptual development (Li et al., 4 Jul 2026). This example is consistent with the larger quantitative results on co-occurrence and ordering.

A concise clarification follows from the example. GRASP is not described as reproducing human text verbatim. Rather, its target is structural fidelity: grouping, discourse role, contrastive positioning, and citation ordering. In that sense, “high-fidelity” refers to the organization of scholarly relations as much as to surface realization.